The logo on the page switched with the mode (#63), but the tab's icon
was always favicon.png, the light logo. web/favicon-dark.png is
logo-dark.svg at 128px, served beside it, and the theme script points
the icon link at whichever matches data-mode, so it follows the theme
the account chose, not only the system. The sign-in page, with no
account, picks by the system's with two media-bound links.
/favicon.ico, which a browser asks for on its own, stays the light one.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Modern's palette was sampled from the 2004 iPodderX icon: a neutral navy,
its screen blue and amber EQ bars. It now takes the new logo's colours,
the dark half from logo-dark.svg (navy ground, #8fc2ea scale, #ff6a1a
needle) and the light half from logo.svg (sky ground, #2f6aa0 scale, the
needle taken down to #c43e00 so white on it clears AA). The pending amber
and the error red moved apart from the needle's orange, and the sign-in
page's copy of the palette follows.
The pages always showed logo.svg, the light variant, even in a dark
theme; logo-dark.svg was never served. It is now, and the app and admin
pages show whichever matches data-mode, dark until the script says light,
as the palette is. The sign-in page, which has no account's theme, picks
by the system's with <picture>.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Production ran four commits past v0.9.0 while the logo's tooltip said
0.9.0, because Cargo.toml's version only moved at a release. It is now
0.9.1-dev, and CLAUDE.md's release steps end by moving to the next -dev
version. No commit hash: the name says there is newer work, and git says
which.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Settings dialog ended with the server's download_dir, read-only and
the same for everyone: it is set in config.toml, so nobody can act on it
from there. The admin page still shows it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The logo sat at the top of the feed list with "iPX" written beside it,
and the page showed the version nowhere. It is now in the top bar just
before the feed buttons, alone, and its tooltip names the app and its
version.
The version is filled in by the server as it sends the page, not by
build.mjs: build.rs reruns only when web/ or package-lock.json changes,
so a release that bumped only Cargo.toml would have kept the page naming
the one before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The needle was #f7931e (#ff9f2e dark), a soft orange close enough in
lightness to the tallest blue bar that the two ran together where they
touch, and weak at favicon size. It is now #ff5500 (#ff6a1a on the dark
background), fully saturated and pushed towards red, away from the bars'
blue. favicon.png and apple-touch-icon.png are re-rendered from logo.svg.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
selectEntry took each read item out of the list the moment you moved on
from it, so the item was not there for the back swipe (or k) to reach: it
went to the one before, or to the list if the item had been first.
Items read while turning from one to the next (a swipe, j and k) now stay
in the list until the reader closes or another item is picked from the
list, and a background refresh keeps them as it keeps the open one.
Picking a row still drops the item left behind at once, as before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
/favicon.png, /apple-touch-icon.png and /logo.svg are kept a day under
fixed names, so after the new logo went out, curl through the tunnel and
browsers still got the old one. The pages now ask for them as
/favicon.png?v=<hash>, the way they already ask for app.js and app.css,
so a changed icon is a new URL for every cache on the way. /favicon.ico
cannot carry a query, as browsers ask for it on their own; it keeps the
day.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The app, the repository (rays/ipx), the image, the compose service and
container, and the data folders on Tower take the new name. What stays:
the 2004 iPodderX and ipodderx-core, which are history; the Postgres
database and login, and ipodderx.sdf1.net with its Access and Authentik
apps, which would each need moving outside this repo; and the first-start
password, as ipx is shorter than the eight characters a password needs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The 2004 icon is a whole radio on a transparent background and not
square, so the tab icon was padded and iOS painted the home-screen icon
on white. web/logo.svg is the radio's screen alone, its tuning scale and
orange needle, laid out as Apple's app icon guidelines ask: opaque and
full-bleed (the system cuts its own corners), a gradient background and
flat foreground layers with hard edges, no highlights or shadows of its
own, nothing thin enough to vanish at 32px. Each layer is a <g>, ready to
split out for Icon Composer. web/logo-dark.svg is the same layers
recoloured.
favicon.png (128) and apple-touch-icon.png (180) are renders of it. The
pages show it from /logo.svg, served outside the auth layer for the
sign-in page, with an app icon's rounded corners. The 2004 icon stays at
/icon.png for anything outside that links to it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On a phone every control was outlined, hairlines ruled off the bars, corners were 7-9px and
dialogs were centred cards: a desktop page, shrunk. Now, below 820px and in every theme, controls
are filled and round (40px circles and capsules, pills for groups), the feed's name is a 26px
large title, the tabs are a segmented control, rows are 44px with 17px text and an inset hairline,
the reader's type is 17px, and dialogs are sheets from the bottom. --edge is the one switch for the
outlines, left on for the high-contrast theme and for prefers-contrast:more. Glass's desktop radii
moved into a min-width query, because at their specificity they outranked the phone's shapes, and
on a phone its #51 rim only catches the top edge: all round a small capsule it read as an outline.
Text boxes were 13.5px, and Safari zooms the page in on focus below 16px, so tapping search zoomed
it; they are 16-17px on a phone now.
Nothing read safe-area-inset-top, so opened from the home screen the top bar, the reader's back
button and the drawer's head sat under the Dynamic Island. --safe-t pads them, and the phone's own
top-bar rule, a shorthand, had also been discarding the side insets for the notch in landscape.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On a phone the reader is fixed over the item list, and the swipe from #49 moved it alone and
faded it to 40%, so the list showed through it and in the strip it uncovered, and again as the
next item slid in from the far side. Now a layer beside the reader, #dpeek, holds what is really
there: the next or previous item, drawn by the same detailHtml the reader uses, with 16px of the
page's background between them; "No more items" past the last; or, swiping right from the
first, a dimmer over the list that lifts as the reader, shadowed along its edge, is drawn off.
On release the swipe carries on from where the finger left it, over 120-250ms by how far is
left, and the neighbour becomes the reader in place; with reduced motion it switches at once.
A touch starting within 14px of the left edge is kept from Safari, which otherwise takes it as
Back and leaves the page mid-swipe. 14px because the back button starts at 16.
Offsets are rounded to whole pixels: at a fractional offset the seam between the reader and its
neighbour drew a stray light line.
The design had the list shift a third of the way across as it is uncovered, as iOS does. #detail
lives inside #shell, and a transform there makes it the containing block for the fixed reader,
so that part is left out; the dimmer and shadow carry it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The glass surfaces had one 1px highlight along the top, which read as flat. They now have a rim
lit from the top-left and a fainter one on the far edges, as box-shadows, and a sheen from the
top-left corner as a background layer. A pseudo-element would have been the usual way, but the
detail pane, file list and dialogs scroll, and an absolutely placed layer in a scroller scrolls
away with the content.
Refraction was looked at and left out again: bending the live page needs backdrop-filter: url(),
which Firefox and Safari lack, and the WebGL libraries (liquid-glass-js, liquidGL) bend a
snapshot of the page that goes stale on a list that scrolls in its own pane.
The sheen lightens a dark panel under its text, so tests/contrast.js now checks every text colour
on a panel under the sheen at full strength. That capped it at 7% in dark mode; 9% put --faint,
--accent and --bad under AA.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A share button in the feed's header, beside an item's "Open the original", and on each file.
It uses the Web Share API where the browser has it, which is the share sheet on a phone, and
copies the link where it does not. A file is shared by the publisher's address, not /media/,
which only someone signed in here can open.
Paid feeds carry the subscriber's access in their address, Patreon's in a token, so sharing one
gives the subscription away. An address that looks like that asks first.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each person has a list for every feed they read and one per feed. An item whose title or text
holds one of the words, matched as whole words so "ai" does not hide everything that "said"
anything, is hidden from them and, since the scanner now keeps each subscriber's filters
separate, is fetched only if someone else still wants it.
Whole-word matching is not something LIKE can do on both SQLite and Postgres, so the matches are
worked out in Rust into a `hidden` table whenever a list changes, someone subscribes, or a scan
brings in new items, and the queries only look that table up. Both new tables are tables rather
than columns because create_missing adds tables but never columns. Hidden counts as read for
the reaper and for "others still want this file", since whoever it is hidden from is as done
with it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
At a flat 60px, a thumb scrolling slightly on the diagonal moved on to the next item by
accident. A swipe now has to cover a quarter of the reader's width, and never less than 100px,
and the reader follows the finger while it is down, so it is plain before letting go whether it
will move on. It slides off on a swipe and the next item slides in from the other side; a short
one springs back. The CHANGELOG also gets back the [0.8.3] link a stray edit had broken.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Glass theme (#43), playback handed to a native shell (#45), and the
bottom bar kept clear of the home indicator (#46), which had no changelog
entry of its own. The unreleased compare link had been left at v0.6.1 since
0.7.0; it points at the new tag now.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
body is a grid of topbar / main / status / player, and nothing in the page
accounted for a display's own intrusions. Mobile Safari hides that by
insetting the layout viewport to the safe area, so the site was fine in a
browser -- but a full-screen shell, the native app or the site added to an
iOS home screen, hands the page the whole display, and the last row landed
under the home indicator with its seek bar and times half cut off.
viewport-fit=cover asks for the whole screen deliberately, and the bars
along the edges now pay for the insets in padding: the bottom for the
indicator, left and right for the notch in landscape. A browser with its
own chrome reports nought and nothing moves.
The padding has to be longhand, and there is a comment saying so, because
the minifier drops the space between a calc() and the value after it in a
shorthand -- padding:7px calc(12px + var(--safe-r))7px ... -- and a browser
then throws the whole declaration away. The bars lost all their padding,
which moved the item list far enough that the pull-to-refresh browser test
stopped finding it; nothing reported an error, and the page still loaded.
buildStyle now fails the build on a calc() run into its neighbour rather
than trusting it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
CarPlay and Android Auto cannot render a web view. Both are template
surfaces, and the only audio they will control is the host's own AVPlayer
or ExoPlayer -- so an app that is "the web UI plus CarPlay" is really "the
web UI whose audio engine is native", and the page had no way to give
playback away.
web/src/native.ts replaces the playback surface of the page's media element
with one that forwards to the host and synthesises the events back. Nothing
in player.ts changes: it only ever speaks to the element, so the player bar,
the row buttons, the EQ bars and the keyboard shortcuts keep working as they
did. Video stays in the page, since CarPlay is audio-only and a native video
layer under a web view buys nothing. In a browser none of it installs.
Position and read are the host's to write. player.ts has been bitten before
by a stale position -- one left paused in another tab saved its older place
over where you had got to -- and a backgrounded web view is exactly that
tab: frozen, holding a time from minutes ago, while the host plays on. So
the beacon becomes a request for the host to save its own clock.
tests/native-bridge.js is what holds the two ends together, and it earned
its place immediately: the src setter called removeAttribute('src'), which
the shim's own override turned into a stop() that switched it back off one
line after enabling it. Silent, and only visible in a car. The stub DOM
moved to tests/dom-stub.js so that test and page-smoke share one harness
rather than two copies.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Translucent panels with backdrop-filter blur and saturation over a soft
coloured wash, light and dark, going solid under prefers-reduced-transparency
and prefers-contrast: more. The sticky filter bar and column headings are
frosted, since the list scrolls under them.
Text on a see-through panel lands on whatever the wash is behind it, so the
palette's hex values alone no longer say whether it clears AA. contrast.js
now samples the wash as the browser composites it on three viewport shapes.
It caught the first light palette at 3.7:1 for faint text, and a tinted
selection that failed everywhere; both were changed.
Left out: SVG displacement-map refraction, which Chromium alone applies to a
backdrop and only on fixed-size shapes.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
tea comment, run without a terminal, waits on stdin and never exits. The example this file gave had the same problem.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
On Tower the host sees the processes inside containers, so the
pkill -x ipx that CLAUDE.md recommended would have killed production's
daemon in the iPodderX container along with the test one. CLAUDE.md now
says to stop a hand-run daemon by its own PID; docs/cli.md keeps
pkill -x for a plain install and warns about the container case.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The proxy sign-in believed Cf-Access-Authenticated-User-Email from any
address in trusted_proxies. On Tower that address is the Docker gateway,
so any container there could name itself anyone (docs/sso.md said as
much, and CLAUDE.md listed it as a known gap).
With [web] access_team and access_aud set, a proxied request must also
carry a Cf-Access-Jwt-Assertion that verifies against Cloudflare's keys
(RS256 only, this application's audience, the team's issuer, not
expired), and the name comes from its email claim. The keys are fetched
at start and again when a token names an unseen key, at most once a
minute, so made-up key ids cannot make every request a request to
Cloudflare. While the keys cannot be had, proxied sign-in is refused;
password and token sign-in are unaffected. Both settings empty, nothing
changes.
jsonwebtoken does the checking, on the aws-lc-rs backend already in the
tree through rustls. Tests sign with throwaway keys in tests/data: a
valid token, another app's audience, expired, a forged signature, HS256,
alg none, the refetch limit, and keys that cannot be fetched. Checked
live on a scratch daemon: the header alone and a forged token got 401,
the admin token still signed in.
vouched_name takes the peer and headers rather than the request: a
&Request held across the new await made the auth middleware's future
unsendable, as a body is not Sync.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Substack puts a post's subtitle in <description> and leaves it out of
content:encoded, and body() took content:encoded alone, so the subtitle
was lost (a known gap in CLAUDE.md). A description is now shown above the
body, as <p><em>, when it is short plain text the body does not already
contain. Podcast feeds that repeat their notes in both, whole or cut short
with an ellipsis, are unchanged; the comparison is by words, since a tag
taken out of the body leaves stray spaces around punctuation.
Checked against Experimental History's feed (subtitles appear) and The
Daily's (notes in both fields, shown once). Entries are inserted with ON
CONFLICT DO NOTHING, so only posts first seen from now on get it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The daemon printed http://<bind>/?token=<token> at every start. The token
signs in as the admin, and in the container that line lands in docker
logs, readable by anyone with Docker access on Tower. It now says where
the token is kept instead; config.toml already has it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Each in light and dark. The published palettes missed AA in 19 places,
mostly Solarized and Catppuccin Latte, so each failing colour is moved
the least distance, toward black or white, that clears every pair it is
drawn in. Solarized dark's base0 had to rise to base1 to read on base02,
so its dim sits between base2 and base1 to keep three steps of type.
tests/contrast.js checks every palette against the pairs the page draws,
and a border that matches the ground it sits on. Its first run caught
Classic's links at 3.7:1 on the source list and Modern's faint at 4.3:1
on inputs; both are tuned. A failed toast moves to --panel, since the
error colours are tuned for the page's grounds, not --raise.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Both have nothing to confirm, so their only button was a lone X at the
foot of the card, below the fold of a long Settings card. A dialog with a
confirm keeps Cancel beside it at the bottom, where the pair belongs.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Measured every theme's palette against WCAG AA. The unread badge's count
on its --accent2 fill fell as low as 2.6:1 (Flat Remix light), and the
unread dot with it; each theme's --accent2 is taken down until white on
it clears 4.5. Tags were --warn or --bad on --raise, short of AA in most
light themes, so they are outlined on the row's own ground instead.
Nordic dark had --line equal to --panel2, so bordered buttons on a
panel2 ground drew no edge at all. Classic's selected row left the
row's icon buttons grey on the blue. A zero badge on a selected row was
--raise on --raise and vanished.
login.html never had a doctype or viewport meta, so it rendered in
quirks mode and at desktop width on a phone, and its light palette sat
under data-theme="light", which nothing ever set. It follows
prefers-color-scheme now, signed out having no account to ask.
Dropped the unused log and users icons. The Classic theme's label is
now just "Classic".
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The scan's events reach everyone, so every browser showed "<feed>: N new" and
"Scanning…" toasts, and refreshed, for everyone's feeds. Now a feed's row, and
its folder's, carries a spinner between feed_start and its done, skip or error;
the list refreshes only for the reader's own feeds; the scan toasts are gone, and
"Downloaded" is said only for a file on screen.
"Check every feed" from the web UI sent a scan of every feed on the server.
Command::Fetch takes an optional `feeds` list -- those feeds and the feeds
inside any OPML among them -- and the web fills it with the asker's
subscriptions. The schedule and the CLI send none, meaning every feed.
Closes#37.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Phase 3 of #18. Two tables: catalogue (each feed's config::Feed as JSON, so a
new feed setting needs no column) and settings (general: the five server
settings the admin page edits). config.toml keeps what is needed before the
database is reached, or decides who gets in: paths, [torrent], [web].
ipx still runs from one in-memory Config, assembled at start from both
(assemble_config). The eight places that saved config.toml and re-read it now
call Ctx::store_cfg, which writes the database and swaps the copy in memory; the
first-run web token, which is config.toml's, is written there.
The first start on a database with no catalogue imports config.toml's feeds and
settings in one transaction whose first insert is the settings row, so two ipx
starting at once cannot both import; it then trims config.toml, keeping the
original as config.toml.pre-database. After that, feeds written into the file are
ignored with a warning. copy-db skips it, and copies both tables.
Rehearsed on a clone of production's database with production's config: all 130
feeds imported, the file trimmed, and the feed list, settings and directory
identical to the live server's.
Postgres connections now ask for no notices. Every CREATE ... IF NOT EXISTS on an
existing table sends one, eleven per open; sqlx logs them, and
tracing-subscriber 0.3.23's per-layer filters then dropped the next line ipx
logged -- the import's own message went missing that way. Proved by toggling it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Sorting by the pin column, or the Pinned tab, was enough. order_sql loses its
pinned_first option, pinning no longer reloads the list, and the tests and
changelog line for #35 go. The NULLS FIRST/LAST ordering from the Postgres work
stays.
Closes#36.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Arcane runs compose in its own container, where /mnt/fast/appdata does not
exist, so an absolute env_file path there failed its update with 'env file not
found'. The file is now ipodderx.env in the content project, referred to
relatively.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Where the database now is and how to reach it, IPX_DATABASE_URL and
IPX_TEST_DATABASE_URL, copy-db, backups, and the title sort on Postgres.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
- IPX_DATABASE_URL (postgres://...) picks the database; unset, it is the SQLite
file as before. Passwords are taken out of anything logged.
- `ipx copy-db <state.db>` copies every table into the empty database the URL
names, in one transaction, and moves the id counters past the copied ids. A
copy of production went across in 14s with every count and column
fingerprint identical.
- With IPX_TEST_DATABASE_URL set, each test gets a Postgres schema of its own;
all 79 pass on both databases. Fixtures write booleans as true/false.
- Sorts say where an item with no value goes (NULLS FIRST going up, LAST going
down): SQLite counts NULL as smallest, Postgres as largest, so "largest first"
on Postgres led with every item that has no file. Tested on both.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
CLAUDE.md and the architecture notes described the SQL schema and migrate(),
both gone: the entities are the schema, create_missing makes what is missing,
and hand-written SQL has to run on SQLite and Postgres both.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The last nineteen functions move to SeaORM: recording feeds, items and
enclosures, managed OPML feeds, folding WordPress's repeated files, and handing a
Patreon creator's files to its shows. Two SQLite-only forms go: GLOB becomes a
LIKE with the underscore escaped (broader, harmlessly: the fold still keys on
`_=` and digits), and UPDATE OR IGNORE becomes an UPDATE ... WHERE NOT EXISTS.
The two transactions are SeaORM transactions.
With nothing left on it, rusqlite goes, with the SQL schema and migrate(). The
entities are the schema: create_missing makes whatever tables and indexes a
database lacks, from them, with CREATE ... IF NOT EXISTS. Production's schema
already has every column migrate() added and none it dropped.
Not SeaORM's schema sync, used until now: despite its docs it drops a unique
index the entities do not describe, so it dropped users_name_lower on every open.
Every `ipx` command then took a write lock, and against a daemon busy writing,
`ipx status` -- the healthcheck -- failed 7 times in 15 where the old code
failed none. Now 15 in 15, as before. On Postgres it would not have started.
WAL is set only when a file is not already in it: setting it takes a lock that
cannot wait out a busy daemon.
Checked on copies of production: a forced scan of all 162 feeds against the real
feeds with no database errors; the feed list, filters, sorts, search and the
reaper's candidates against the old code on the same data, earlier in the
branch. The column comments from the SQL schema move to the entities.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Twelve enclosure functions move to SeaORM: recording, the download queue,
marking done or failed, requeueing, and what the reaper may delete. INSERT OR
IGNORE becomes ON CONFLICT DO NOTHING; the reaper's read verdict is true or
false rather than 1 or 0, which Postgres would type as a 32-bit integer and
refuse to read as an i64; `read = 1` and `flagged = 1` test the booleans
themselves. retention::run and its callers (reap, rm, retire_group,
retire_stranded) become async.
The reaper deletes files, so it was checked on a copy of production against the
old SQL on the same file: all 2,195 candidates, identical and in the same order.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
